1. King Fahd University of Petroleum and Minerals CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua A High Throughput Network-on-Chip Architecture for System-on-Chip Interconnect Abdelhafid Bouhraoua and M.E.S El-Rabaa Computer Engineering Department (COE) College of Computer Science and Engineering (CCSE) King Fahd University of Petroleum and Minerals (KFUPM) Dhahran, Eastern Province, Saudi Arabia

2. King Fahd University of Petroleum and Minerals 2 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Outline Overview and Motivation Fat Tree Network Properties Modified Fat Tree Router Architecture Performance Evaluation Conclusion

3. King Fahd University of Petroleum and Minerals 3 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Outline Overview and Motivation Fat Tree Network Properties Modified Fat Tree Router Architecture Performance Evaluation Conclusion

4. King Fahd University of Petroleum and Minerals 4 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Networks-on-Chips “Route Packets NOT Wires”, William J. Dally Idea: Build a Complete on-Chip Network Unified Communication Model (Similar to OSI Stack) No Ad-hoc Effort Standardized Interfacing (May be provided by IP Vendors) Unified Network Elements (Routers, Link Interfaces) No Design required by the SoC Teams Flexible Interconnect and Reduced Global Wiring

5. King Fahd University of Petroleum and Minerals 5 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua NoC Requirements Performance How fast packets are moved across the network? How much traffic is carried at the same time and for how long? Overhead How Big is its required Size (in Gates) ? Adaptivity Does it Adapt Easily to new Designs ? Complexity How Easy is Interfacing to it ?

6. King Fahd University of Petroleum and Minerals 6 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Previous Work Majority directly derived from other research (Interconnection Networks for Parallel Architectures) Router architectures directly derived from inter-chip architectures where the routers were implemented on a single chip  substantial overhead. Focus on the router architecture alone to achieve certain goals in latency Circuit switching techniques introduced to provide a certain guarantee for the latency. Added complexity to achieve guaranteed latency is an overkill in the on-chip context. Did not fully take advantage of the fact that the network is on-chip where the main gain is no-pin limitation.

7. King Fahd University of Petroleum and Minerals 7 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Which Network? Most Straightforward  Crossbar Good Throughput (maxes at 66%) Non Scalable (Quadratic) Complexity Of Implementation for Higher Number of I/Os.

8. King Fahd University of Petroleum and Minerals 8 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua 2-D Mesh Very Popular Topology in NoCs. Very Suitable for the 2D nature of Chip Floorplanning (Tiling) Very High Constraints Inefficient routing algorithms (deadlock-free by construction) Efficient routing algorithms (Complex implementation) Poor performance: Saturation reached at 30 %.

9. King Fahd University of Petroleum and Minerals 9 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Analysis Low throughput. Means: latency cannot be guaranteed above the maximum throughput levels Low throughput cause by contention over the output ports of routers among several incoming packets Cannot prevent contention from happening. Contention makes router architectures more complex because they need to integrate buffering and prioritization logic. Routers that implement both packet and circuit switching makes the architecture even more complex.

10. King Fahd University of Petroleum and Minerals 10 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Methodology Take advantage of the On-Chip Context: Design frozen before tape out No internal IO limitations Aim for a High Throughput Architecture Circuitry used at 30% of its maximum is NOT an optimal Solution (Clock frequency, power). Reduced router size Integrate a large number of routers Wormhole routing vs. Store and Forward Reduce required buffers in routers

11. King Fahd University of Petroleum and Minerals 11 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Fat Tree RRRR RRRR RRRR CCCCCCCC Bidirectional multistage or folded multistage networks Bidirectional multistage are two entities: The Fat Tree (FT) The butterfly. Fat Tree better than butterfly (previous work) What topology resembles a crossbar? Banyans or Multistage Interconnection Networks. n+1 Stages (or rows) Size is Routers = n x 2 n Clients = 2 n+1 Diameter = 2logk + 1; n = log k

12. King Fahd University of Petroleum and Minerals 12 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Outline Overview and Motivation Fat Tree Network Properties Modified Fat Tree Router Architecture Performance Evaluation Conclusion

13. King Fahd University of Petroleum and Minerals 13 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Routing in Fat Tree Routing reduced to routing in a binary tree. Binary Trees Three Routing Directions UP RIGHT LEFT Router UP LEFT RIGHT

14. King Fahd University of Petroleum and Minerals 14 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Routing in Fat Tree Router (r,c) [0, l[ U [u, 2 n+1 -1] [l, l+2 r-1 [[l+2 r-1, u[ Lower bound l : smallest address reached from the router (r,c). Smallest address within the range obtained by clearing the lowest r bits of the column c. l = (c/2 r ) x 2 r. Upper bound u: largest address reached from the router (r,c). Largest address obtained by adding 2 r to the lower bound l. u = l+ 2 r. Matrix n rows x 2 (n-1) columns. Router (r,c) r : row index (rows are indexed from 0 to n-1) c: column index (columns are indexed from 0 to 2 (n+1) -1) Size of the clients’ address space reachable using the downside ports is equal to 2 r It is always a continuous interval of addresses of the form [l, u[.

15. King Fahd University of Petroleum and Minerals 15 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Routing In Fat Tree RRRR RRRR RRRR CCCCCCCC “Summit” Routers Routing UP: Adaptive Routing Down: Deterministic Alternate Paths

16. King Fahd University of Petroleum and Minerals 16 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Outline Overview and Motivation Fat Tree Network Properties Modified Fat Tree Router Architecture Performance Evaluation Conclusion

17. King Fahd University of Petroleum and Minerals 17 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Contention in Fat Tree Packets coming from the UP links are never routed up Only packets coming from the bottom links are routed up. Since the number of UP links is equal to the number of bottom links, there cannot be any contention when routing up. Contention occurs only when going down. Bottom links are split in RIGHT and LEFT links, deterministic routing of packets will lead to contention. UP LEFTRIGHT Many Choices for Going UP Contention on the way down

18. King Fahd University of Petroleum and Minerals 18 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Modified Fat Tree Doubling of downward links eliminates contention

19. King Fahd University of Petroleum and Minerals 19 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Outline Overview and Motivation Fat Tree Network Properties Modified Fat Tree Router Architecture Performance Evaluation Conclusion

20. King Fahd University of Petroleum and Minerals 20 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Router Architecture No Crossbar No Buffers (Pushed to the Clients) Every downstream input simultaneously connected to two outputs. Contention eliminated between the inputs going downstream. Number of outputs is 2k+2 for k inputs (case of when the router is a summit) Router models differ from each other only by two items: Number of input and output ports on the down link Routing function constants (r,c)

21. King Fahd University of Petroleum and Minerals 21 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Routing Circuitry All network elements are constants and frozen at design time. All lower bound and upper bound values, used to generate the routing functions, are constants for each router. These constants are entered as inputs into the routing function Routing Function implemented using comparators. Constants needed by the routing function are: l L = l + 2 r -1 u Address ≥ <≥ < ABAB ABAB ABAB l L u ≥ <≥ < ≥ <≥ < LEF T RIGH T UP

22. King Fahd University of Petroleum and Minerals 22 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Client Interface Buffers pushed to the Client Interfaces Each incoming link is terminated with a FIFO memory. The different FIFO memories connected to the client through a single shared bus. Client/IP Block FIFO Down Links (from router) Up Link FIFO Bus can be wider to perform data transfers faster than what is received in the FIFOs. The size of FIFOs customizable by design team according to the specifications

23. King Fahd University of Petroleum and Minerals 23 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Outline Overview and Motivation Fat Tree Network Properties Modified Fat Tree Router Architecture Performance Evaluation Conclusion

24. King Fahd University of Petroleum and Minerals 24 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Simulation Conditions Uniform Traffic Generation Uniform Distribution of Destinations Traffic Rate constant fraction of Maximum Link Bandwidth Variable Packet Size (within a predetermined range; eg. 64 bytes +/- 10%) Simulation Platform: Cycle-based C-based. Developed for this purpose

25. King Fahd University of Petroleum and Minerals 25 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Throughput More than 90% Throughput achieved Compare with Regular Fat Tree

26. King Fahd University of Petroleum and Minerals 26 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Latency Latency has linear progression Large component of Latency spent in Receiver’s FIFOs

27. King Fahd University of Petroleum and Minerals 27 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Area and Speed Buffer-less architecture less costly

28. King Fahd University of Petroleum and Minerals 28 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Client Buffer Utilization Buffers pushed to the client interfaces. Considerable number of buffer lanes is necessary for every client interface. Simulations shows a linear progression of the maximum number of lanes used during operation. Obtained figures are an order of magnitude lower than the number required by the architecture. Number of buffer lanes in the client interface can be tailored to suit the class of applications at hand while reducing buffering area.

29. King Fahd University of Petroleum and Minerals 29 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Outline Overview and Motivation Fat Tree Network Properties Modified Fat Tree Router Architecture Performance Evaluation Conclusion

30. King Fahd University of Petroleum and Minerals 30 CCSE – COESOC 2006 – Tampere, 14-16 Nov. 2006Abdelhafid Bouhraoua Conclusion A contention-free modified FT architecture is proposed. Proposed architecture achieves maximum theoretical throughput and has smaller latency than conventional FTs. Latency increases linearly with input load. Achieved performance is actual performance using a contention-free network. The area of the network is kept small because of the absence of buffers in the router architecture. Number of buffer lanes in the client interfaces can be tailored for a specific platform to suit the class of applications at hand while reducing buffering area.